Inspecting Things Instead of People

A large share of the genuinely successful deployments in this field point cameras at objects rather than at people. They are worth listing, because they are frequently overlooked in favour of applications that are harder and more contentious.

Infrastructure condition

Road surface defects from vehicle-mounted cameras: cracking, potholes, marking wear.

Rail infrastructure from inspection vehicles: fastenings, sleepers, clearance.

Bridges and structures from drone imagery: cracking, spalling, corrosion.

Utility assets: insulator damage, vegetation encroachment on lines, corrosion on towers.

Why it works: the defect is visual by nature, the alternative is a person with binoculars, and coverage rather than precision is the value.

Agriculture and environment

Crop condition and disease from ground or aerial imagery.

Weed identification for targeted treatment, which reduces chemical use measurably.

Livestock counting and condition, which is welfare-relevant and uncontentious.

Wildlife monitoring from camera traps, where the alternative is a person reviewing hundreds of thousands of images.

Land use change from satellite and aerial imagery.

These have a strong record and the error tolerance is usually generous: a system that finds most of the affected area is useful even at moderate accuracy.

Industrial assets

Corrosion, leaks and wear on plant equipment.

Gauge and dial reading where legacy equipment has no digital output — an unglamorous application that saves real manual rounds.

Thermal imaging for hot spots, which is a different sensor and the same analysis.

Conveyor and belt condition.

Stock levels in bins and yards, where a count is the whole answer.

Why these deserve more attention

No people in frame, or people incidental, which removes the privacy analysis, the fairness question and most of the regulatory burden.

Ground truth is obtainable. You can go and look at the bridge.

Errors are recoverable. A missed crack is found on the next pass; a wrongly flagged one costs an inspection.

The comparison is against sampling. Manual inspection covers a fraction of the asset; a camera on a vehicle covers all of it. Even a mediocre model beats not looking.

What still needs care

Coverage claims. A system that inspects everything at moderate accuracy is not the same as inspecting a sample carefully, and reporting it as equivalent overstates it.

Severity assessment. Detecting a crack is much easier than judging whether it matters, and the second is an engineering judgement that should stay with an engineer.

Ground truth cost. Verifying findings requires site visits, and the evaluation budget is frequently omitted.

Environmental variation: wet roads, low sun, seasonal vegetation.

Where people appear incidentally

Vehicle-mounted road inspection captures pedestrians and number plates.

Drone survey captures gardens and windows.

This is incidental collection and it still requires handling: blur or discard at source, retain only the detections, and state the retention.

Processing at the edge, discarding frames and keeping only the finding removes almost the entire concern and is technically straightforward for these applications.

Handling incidental people

Object inspection captures people anyway, and the handling is straightforward.

Blur or mask at capture, on the device, before anything is stored.

Retain the detection, not the frame, which for defect finding is all that is needed.

Number plates and house frontages need the same treatment on vehicle-mounted surveys.

State it in the notice for the survey activity.

This removes almost the entire concern for a class of applications and costs a configuration setting.

Coverage against precision

The trade that makes these applications valuable, and it should be stated honestly.

Manual inspection examines a sample carefully.

A camera on a vehicle examines everything, less carefully.

Complete coverage at moderate accuracy frequently beats a careful sample, because defects are not uniformly distributed and a sample misses whole areas.

Say which you are providing. Reporting automated survey as equivalent to careful inspection overstates it.

The useful framing: this finds candidates across the whole asset; an engineer assesses severity. Separating detection from judgement is what makes it defensible.